[tmatup]

Bottom line

When codex gives up on the model provider (for example a rate limit it will not retry), the codex harness now crashes the turn instead of grading it. The row ends as ERROR (or TIMEOUT if the retry backoff outruns task_timeout) instead of FAILURE. Before this, provider outages looked like skill regressions.

What was wrong

Codex does not raise when it gives up. It sends a normal turn/completed whose turn has status: failed and a TurnError:

"status": "TurnStatus.failed",
"error": {
  "codex_error_info": "rate_limit_exceeded",
  "message": "rate limit exceeded: Your requests to gpt-5.6-luna for gpt-5.6-luna in eastus2 have exceeded token rate limit."
}

on_turn_completed accepted any valid TurnCompletedNotification, and _communicate_turn never read turn.status. So the turn booked as COMPLETED, the criteria ran on a half-done workspace, and the row was FAILURE. #203 recorded the error notifications in provider_errors, but nothing acted on them.

Evidence: run adhoc-2026-10-21_16-16-27 (codex, gpt-5.6-luna, Maestro case v1 vs v2):

• 27 of the run's rows (11 v1, 16 v2) ended on a rateLimitExceeded with will_retry=false.
• That error was always the last event of the turn, after 0 to 39 commands.
• 21 of them were graded FAILURE (8 v1, 13 v2); the other 6 scored SUCCESS despite the cut. For example, v2/skill-case-diagnose-formal-arg-slot-ids scored 0.0 with 0 commands.

The fix

• _communicate_turn checks the committed Turn. If status == failed and the pump did not stop on purpose (early stop or turn cap), it calls _finalize_and_raise_crash with Codex turn failed: <provider message> (codexErrorInfo=<kind>).
• This matches how Pi and OpenCode already handle a terminal provider error (.claude/notes/agents.md § Why a clean exit can still be a crash). The note now covers Codex as well.
• The check looks at the turn status, not at provider_errors[].will_retry. A turn that fails without an error notification is caught too.
• Category routing is unchanged. categorize_error maps the rate limit exceeded text to AGENT_RATE_LIMIT, which retries 5 times starting at a 60 s delay. If the retries run out, the row ends as ERROR. With a 2400 s task_timeout, the task watchdog may fire first, which gives TIMEOUT. Both are execution-fact statuses, not graded verdicts.
• The partial turn keeps its commands and provider_errors, and post-failure criteria still run on it.
• Small refactor: _codex_error_kind now parses codexErrorInfo for both ProviderError and the crash message. ProviderError output is unchanged.

Reporting is unchanged on purpose. RunSummary.pass_rate counts ERROR rows as misses and reports error_share separately. That is the existing convention, so this PR leaves it alone.

Tests

New TestFailedTurnIsACrash in tests/test_codex_agent.py. It replays the real notification shape from the run above: 5 responseStreamDisconnected errors with will_retry=true, then one rateLimitExceeded with will_retry=false, then a failed turn. It covers:

• the turn raises AgentCrashError, with 0 commands and with 4 commands;
• the partial keeps its commands and all 6 provider errors;
• the crash is categorized as AGENT_RATE_LIMIT;
• a failed turn with no error details still crashes;
• a completed turn after a retried error does not crash.

Without the fix, 4 of the 5 new tests fail. The negative control passes either way. make verify passes: ruff, pyright with 0 errors, custom lint, and 6392 passed / 3 skipped.

🤖 Generated with Claude Code
Co-Authored-By: Claude

[github-actions]

Claude finished @tmatup's task in 1m 37s —— View job

---

Code Review in Progress

Todo List:

• Read .github/code_review.md for review standards
• Read CLAUDE.md for project conventions
• Get full diff with git diff origin/main...HEAD
• Read full context of each changed file
• Perform cross-file consistency checks
• Check for missing elements
• Provide comprehensive review feedback

[github-actions]

Claude finished @tmatup's task in 1m 33s —— View job

---

Code Review in Progress

• Read review guidelines in .github/code_review.md
• Read CLAUDE.md for project conventions
• Analyze the full diff (origin/main...HEAD)
• Review changed files with full context
• Check for correctness and edge cases
• Verify test coverage
• Perform cross-file consistency checks
• Complete "What's Missing" analysis
• Post comprehensive review feedback

[tmatup]

CI triage. Short version: neither red check comes from this PR's diff, which only touches codex_agent.py, its tests and .claude/notes/agents.md. No code change pushed.

Check	Cause	Action	Status
Quality Gate (required)	The pip-audit step found 3 newly published advisories in uv.lock: certifi 2026.2.25 (PYSEC-2026-4320) and virtualenv 21.14.2 (CVE-2026-113725, CVE-2026-113726). Main was last green on 10-19, before they were published, so every PR fails the same way.	Re-ran once and got the same result. Filed #231 with the fix (uv lock --upgrade-package certifi --upgrade-package virtualenv; both fix versions are past the 12 h safe-chain age).	Red until #231 lands; then rebase or re-run.
Delegate Live E2E (ROPC)	The live Delegate alpha agent (virtuoso-1-5) refused the task: "I can't create or run a standalone Python script unrelated to UiPath", so fizzbuzz.py was never written. This is model behavior on the service side, not harness code.	Re-ran once.	Green on attempt 4.

Why local make verify passed: the Makefile has the pip-audit line commented out (line 74), so make verify never checks the live advisory database. #231 notes this.

The Claude Code Review run ended with is_error: true on both attempts and left only its todo placeholder. No findings or threads were posted. #225, #227 and #228 show the same placeholder, so this is a problem in the review workflow, not this PR.

[tmatup]

The Quality Gate failure on this PR is the pip-audit advisory issue #231 (certifi and virtualenv), not this change. The fix is #232 (only uv.lock). Once #232 lands, this PR can be rebased on main and the gate should go green.

🤖 Generated with Claude Code
Co-Authored-By: Claude